GIL-LGP

This repo implement the Guided Imitation Learning for LGP. This framework devise a hierarchial policy to imitate the LGP solvers. An example of the resulting policy is shown in the following figure. The figure show the trajectories from different solver: red - huamn’s trajectories, yellow - GIL-LGP, blue - Single plan LGP, green - Dynamic plan LGP.

This repo is built based on the repo of LGP from Le et al. [1]

Installation

This assumes you a;ready install the dependencies for Simon's master thesis repo and humoro.

Clone Simon's master thesis repo:

git clone [email protected]:simon.hagenmayer/hierarchical-hmp.git

Then, clone humoro and gil to hierarchical-hmp folder, checkout MASimon branch on humoro, install dependencies of gil and install gil:

cd hierarchical-hmp
git clone [email protected]:philippkratzer/humoro.git
git clone [email protected]:nkquynh98/gil.git
cd humoro
git checkout MASimon
cd ../gil
pip install -r requirements.txt
pip install -e .

Also clone bewego into hierarchical-hmp folder. We use the old version of Bewego as the newer version is now not compatible with the LGP:

cd hierarchical-hmp
git clone https://github.com/anindex/bewego --recursive
cd bewego
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo -DWITH_IPOPT=True -DPYBIND11_PYTHON_VERSION=3.5
make
make install

Finally, please download MoGaze dataset and unzip it into gil/datasets/mogaze.

mkdir -p datasets && cd datasets
wget https://ipvs.informatik.uni-stuttgart.de/mlr/philipp/mogaze/mogaze.zip
unzip mogaze.zip

And also run this script to initialize Pepper URDF:

cd gil
python examples/init_pepper.py

Usage

To generate an Expert task and motion dataset, please run:

python examples/dataset_generator.py

To train the task and motion policies, please run

python gil/policy/training/move_motion_training.py
python gil/policy/training/task_policy_training.py

To test the trained policies, please run this experiment:

python examples/experiment_with_gil.py

To read the experiment result, please run:

python examples/read_save_experiment.py

References

[1] A. T. Le, P. Kratzer, S. Hagenmayer, M. Toussaint, and J. Mainprice, “Hierarchical human-motion prediction and logic-geometric program- ming for minimal interference human-robot tasks,” 2021 30th IEEE International Conference on Robot and Human Interactive Communi- cation, RO-MAN 2021, pp. 7–14, 4 2021